radv: Move some helper functions to the radv_shader.h header file.
authorTimur Kristóf <timur.kristof@gmail.com>
Mon, 24 Feb 2020 14:44:48 +0000 (15:44 +0100)
committerMarge Bot <eric+marge@anholt.net>
Wed, 11 Mar 2020 08:34:10 +0000 (08:34 +0000)
Move calculate_tess_lds_size and get_tcs_num_patches to radv_shader.h
ACO will need to call these functions too.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3964>

src/amd/vulkan/radv_nir_to_llvm.c
src/amd/vulkan/radv_shader.h

index 5bce7b156c38eede5e28b2d843426d5478782457..81b610ab137648ad7976034c13167daf94f670df 100644 (file)
@@ -117,87 +117,6 @@ static LLVMValueRef get_rel_patch_id(struct radv_shader_context *ctx)
        }
 }
 
-static unsigned
-get_tcs_num_patches(struct radv_shader_context *ctx)
-{
-       unsigned num_tcs_input_cp = ctx->args->options->key.tcs.input_vertices;
-       unsigned num_tcs_output_cp = ctx->shader->info.tess.tcs_vertices_out;
-       uint32_t input_vertex_size = ctx->tcs_num_inputs * 16;
-       uint32_t input_patch_size = ctx->args->options->key.tcs.input_vertices * input_vertex_size;
-       uint32_t num_tcs_outputs = util_last_bit64(ctx->args->shader_info->tcs.outputs_written);
-       uint32_t num_tcs_patch_outputs = util_last_bit64(ctx->args->shader_info->tcs.patch_outputs_written);
-       uint32_t output_vertex_size = num_tcs_outputs * 16;
-       uint32_t pervertex_output_patch_size = ctx->shader->info.tess.tcs_vertices_out * output_vertex_size;
-       uint32_t output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
-       unsigned num_patches;
-       unsigned hardware_lds_size;
-
-       /* Ensure that we only need one wave per SIMD so we don't need to check
-        * resource usage. Also ensures that the number of tcs in and out
-        * vertices per threadgroup are at most 256.
-        */
-       num_patches = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp) * 4;
-       /* Make sure that the data fits in LDS. This assumes the shaders only
-        * use LDS for the inputs and outputs.
-        */
-       hardware_lds_size = 32768;
-
-       /* Looks like STONEY hangs if we use more than 32 KiB LDS in a single
-        * threadgroup, even though there is more than 32 KiB LDS.
-        *
-        * Test: dEQP-VK.tessellation.shader_input_output.barrier
-        */
-       if (ctx->args->options->chip_class >= GFX7 && ctx->args->options->family != CHIP_STONEY)
-               hardware_lds_size = 65536;
-
-       num_patches = MIN2(num_patches, hardware_lds_size / (input_patch_size + output_patch_size));
-       /* Make sure the output data fits in the offchip buffer */
-       num_patches = MIN2(num_patches, (ctx->args->options->tess_offchip_block_dw_size * 4) / output_patch_size);
-       /* Not necessary for correctness, but improves performance. The
-        * specific value is taken from the proprietary driver.
-        */
-       num_patches = MIN2(num_patches, 40);
-
-       /* GFX6 bug workaround - limit LS-HS threadgroups to only one wave. */
-       if (ctx->args->options->chip_class == GFX6) {
-               unsigned one_wave = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp);
-               num_patches = MIN2(num_patches, one_wave);
-       }
-       return num_patches;
-}
-
-static unsigned
-calculate_tess_lds_size(struct radv_shader_context *ctx)
-{
-       unsigned num_tcs_input_cp = ctx->args->options->key.tcs.input_vertices;
-       unsigned num_tcs_output_cp;
-       unsigned num_tcs_outputs, num_tcs_patch_outputs;
-       unsigned input_vertex_size, output_vertex_size;
-       unsigned input_patch_size, output_patch_size;
-       unsigned pervertex_output_patch_size;
-       unsigned output_patch0_offset;
-       unsigned num_patches;
-       unsigned lds_size;
-
-       num_tcs_output_cp = ctx->shader->info.tess.tcs_vertices_out;
-       num_tcs_outputs = util_last_bit64(ctx->args->shader_info->tcs.outputs_written);
-       num_tcs_patch_outputs = util_last_bit64(ctx->args->shader_info->tcs.patch_outputs_written);
-
-       input_vertex_size = ctx->tcs_num_inputs * 16;
-       output_vertex_size = num_tcs_outputs * 16;
-
-       input_patch_size = num_tcs_input_cp * input_vertex_size;
-
-       pervertex_output_patch_size = num_tcs_output_cp * output_vertex_size;
-       output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
-
-       num_patches = ctx->tcs_num_patches;
-       output_patch0_offset = input_patch_size * num_patches;
-
-       lds_size = output_patch0_offset + output_patch_size * num_patches;
-       return lds_size;
-}
-
 /* Tessellation shaders pass outputs to the next shader using LDS.
  *
  * LS outputs = TCS inputs
@@ -4123,7 +4042,16 @@ LLVMModuleRef ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm,
                                ctx.tcs_num_inputs = args->options->key.tcs.num_inputs;
                        else
                                ctx.tcs_num_inputs = util_last_bit64(args->shader_info->vs.ls_outputs_written);
-                       ctx.tcs_num_patches = get_tcs_num_patches(&ctx);
+                       ctx.tcs_num_patches =
+                               get_tcs_num_patches(
+                                       ctx.args->options->key.tcs.input_vertices,
+                                       ctx.shader->info.tess.tcs_vertices_out,
+                                       ctx.tcs_num_inputs,
+                                       ctx.args->shader_info->tcs.outputs_written,
+                                       ctx.args->shader_info->tcs.patch_outputs_written,
+                                       ctx.args->options->tess_offchip_block_dw_size,
+                                       ctx.args->options->chip_class,
+                                       ctx.args->options->family);
                } else if (shaders[i]->info.stage == MESA_SHADER_TESS_EVAL) {
                        ctx.abi.load_tess_varyings = load_tes_input;
                        ctx.abi.load_tess_coord = load_tess_coord;
@@ -4225,7 +4153,14 @@ LLVMModuleRef ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm,
 
                if (shaders[i]->info.stage == MESA_SHADER_TESS_CTRL) {
                        args->shader_info->tcs.num_patches = ctx.tcs_num_patches;
-                       args->shader_info->tcs.lds_size = calculate_tess_lds_size(&ctx);
+                       args->shader_info->tcs.lds_size =
+                               calculate_tess_lds_size(
+                                       ctx.args->options->key.tcs.input_vertices,
+                                       ctx.shader->info.tess.tcs_vertices_out,
+                                       ctx.tcs_num_inputs,
+                                       ctx.tcs_num_patches,
+                                       ctx.args->shader_info->tcs.outputs_written,
+                                       ctx.args->shader_info->tcs.patch_outputs_written);
                }
        }
 
index b38710e6fcfafee4dcc0548708c2592a4cb44e0c..30445353e7cc99b6eac53a62b8b61cb98649d7dd 100644 (file)
@@ -505,6 +505,82 @@ shader_io_get_unique_index(gl_varying_slot slot)
        unreachable("illegal slot in get unique index\n");
 }
 
+static inline unsigned
+calculate_tess_lds_size(unsigned tcs_num_input_vertices,
+                       unsigned tcs_num_output_vertices,
+                       unsigned tcs_num_inputs,
+                       unsigned tcs_num_patches,
+                       unsigned tcs_outputs_written,
+                       unsigned tcs_per_patch_outputs_written)
+{
+       unsigned num_tcs_outputs = util_last_bit64(tcs_outputs_written);
+       unsigned num_tcs_patch_outputs = util_last_bit64(tcs_per_patch_outputs_written);
+
+       unsigned input_vertex_size = tcs_num_inputs * 16;
+       unsigned output_vertex_size = num_tcs_outputs * 16;
+
+       unsigned input_patch_size = tcs_num_input_vertices * input_vertex_size;
+
+       unsigned pervertex_output_patch_size = tcs_num_output_vertices * output_vertex_size;
+       unsigned output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
+
+       unsigned output_patch0_offset = input_patch_size * tcs_num_patches;
+
+       return output_patch0_offset + output_patch_size * tcs_num_patches;
+}
+
+static inline unsigned
+get_tcs_num_patches(unsigned tcs_num_input_vertices,
+                       unsigned tcs_num_output_vertices,
+                       unsigned tcs_num_inputs,
+                       unsigned tcs_outputs_written,
+                       unsigned tcs_per_patch_outputs_written,
+                       unsigned tess_offchip_block_dw_size,
+                       enum chip_class chip_class,
+                       enum radeon_family family)
+{
+       uint32_t input_vertex_size = tcs_num_inputs * 16;
+       uint32_t input_patch_size = tcs_num_input_vertices * input_vertex_size;
+       uint32_t num_tcs_outputs = util_last_bit64(tcs_outputs_written);
+       uint32_t num_tcs_patch_outputs = util_last_bit64(tcs_per_patch_outputs_written);
+       uint32_t output_vertex_size = num_tcs_outputs * 16;
+       uint32_t pervertex_output_patch_size = tcs_num_output_vertices * output_vertex_size;
+       uint32_t output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
+
+       /* Ensure that we only need one wave per SIMD so we don't need to check
+        * resource usage. Also ensures that the number of tcs in and out
+        * vertices per threadgroup are at most 256.
+        */
+       unsigned num_patches = 64 / MAX2(tcs_num_input_vertices, tcs_num_output_vertices) * 4;
+       /* Make sure that the data fits in LDS. This assumes the shaders only
+        * use LDS for the inputs and outputs.
+        */
+       unsigned hardware_lds_size = 32768;
+
+       /* Looks like STONEY hangs if we use more than 32 KiB LDS in a single
+        * threadgroup, even though there is more than 32 KiB LDS.
+        *
+        * Test: dEQP-VK.tessellation.shader_input_output.barrier
+        */
+       if (chip_class >= GFX7 && family != CHIP_STONEY)
+               hardware_lds_size = 65536;
+
+       num_patches = MIN2(num_patches, hardware_lds_size / (input_patch_size + output_patch_size));
+       /* Make sure the output data fits in the offchip buffer */
+       num_patches = MIN2(num_patches, (tess_offchip_block_dw_size * 4) / output_patch_size);
+       /* Not necessary for correctness, but improves performance. The
+        * specific value is taken from the proprietary driver.
+        */
+       num_patches = MIN2(num_patches, 40);
+
+       /* GFX6 bug workaround - limit LS-HS threadgroups to only one wave. */
+       if (chip_class == GFX6) {
+               unsigned one_wave = 64 / MAX2(tcs_num_input_vertices, tcs_num_output_vertices);
+               num_patches = MIN2(num_patches, one_wave);
+       }
+       return num_patches;
+}
+
 void
 radv_lower_fs_io(nir_shader *nir);